Bladder cancer segmentation on multispectral images

Nonmuscle Invasive Bladder Cancer (NMIBC) has high incidence, and close follow-up with cystoscopy is necessary due to its high recurrence rate after initial treatment, estimated to be as high as 75%. Because of the high recurrence rate, it is vital that the detection of bladder cancer is improved. Computer automated detection algorithms have shown to be exceptionally effective in achieving this goal. This paper presents the first automated segmentation algorithm for bladder cancer in endoscopic images. The second purpose of this study is to determine which modality is best suited for computer-aided segmentation of bladder cancer. Gabor and color features are extracted from 20 patients in four different modalities with a block-based strategy. Three different classifiers are used to classify the blocks and post-processing is applied to obtain a segmented region. The best classification results were obtained by using a support vector machine and the Spectrum B modality. Additionally, color features were found to be effective for obtaining segmentations comparable to experts.